FreqMamba: Spatial-Frequency Fusion and State Space Sequence Modeling for Deepfake Detection

The rapid evolution of deepfake generation techniques has made high-fidelity facial manipulation a critical threat to social credibility and personal privacy, demanding detection algorithms with strong cross-domain generalization. Existing methods suffer from two fundamental limitations: spatial-domain approaches cannot capture imperceptible forgery artifacts, while frequency-aware methods lack effective integration of spatial semantic and spectral features. To address these challenges, we propose FreqMamba, an end-to-end face forgery detection framework that adaptively aggregates spatial semantic features and frequency-domain artifacts via a gated late-fusion mechanism, and performs global sequence modeling using a bidirectional vision state space model (Vim). FreqMamba consists of three core components: a CNN branch for spatial semantic features, a hierarchical discrete wavelet transform (DWT) branch for fine-grained frequency artifacts, and a bidirectional Mamba backbone for global sequence modeling with linear complexity. The gated fusion mechanism adaptively combines multi-branch features, enhancing responses in forgery-rich regions while suppressing irrelevant noise. Trained exclusively on FaceForensics++ (c23), FreqMamba achieves strong cross-domain performance: on Celeb-DF v2, it attains 0.7767 AUC, surpassing a comparable-parameter CNN baseline (1.14 M parameters, 0.7262 AUC) by 5.05 percentage points; on the real-world WildDeepfake dataset, it achieves 0.6993 AUC, outperforming the lightweight CNN baseline (0.6272 AUC) by 7.21 points. Ablation studies confirm that DWT frequency priors and Mamba sequence modeling exhibit synergistic effects, and Grad-CAM visualizations validate the model's focus on critical forgery regions. FreqMamba provides an effective approach for generalized face forgery detection in cross-domain scenarios.